Professional Documents
Culture Documents
Ijmqe 150013
Ijmqe 150013
Ijmqe 150013
6, 206 (2015)
c EDP Sciences 2015
DOI: 10.1051/ijmqe/2015013
Abstract. This article proposes a mathematical procedure designed to generate an economical measure-
ment model for determining the risk of decision error (customer risk). The model includes the guardbanding
to reduce the impacts of measurement errors. A mathematical model is provided as an example, and con-
clusions are drawn.
Fig. 1. Measurement errors with and without guard band. (a) False acceptance error with and without guard band. (b) False
rejection error with and without guard band.
To meet customer requirements and to avoid the high ω = U − εU . Hereafter, υ and ω are referred to as the
cost of passing bad product to customers, the consumer’s lower and upper inspection limits, respectively. Since the
risk should not exceed a specified value. In this case, the conformance of a product is determined on the basis of
guardbanding υ and ω are set inside the specification lim- repeated measurements, a product passes the inspection
its L and U . Let υ = L + εL and ω = U − εU , then εL and is shipped to the customer if y [υ, ω] and RC = β
and εU are positive. On the other hand, if the consumer’s represent the customer risk (the risk of delivering a prod-
risk exceeds the specified value, then εL or εU may be uct that is intrinsically unacceptable but that is accepted
negative. In this paper, we focus on meeting customer’s by the control means), i.e., y [υ, ω] and X[L, U ].
requirements so that εL and εU are positive. In general, The expected cost by falsely accepting a defect, de-
the consumer’s risk should be much lower than the pro- noted by β , is the conditional probability that a prod-
ducer’s risk because the cost of letting bad product get uct will be accepted given that it is defective. it is then
to consumers is usually much higher than the cost of re- given by:
jecting good product. The difference between the prod- β
β = , (1)
uct tolerance and the length of the guardbanding interval 1 − P (X ∈ [L, U ])
is (U − L) − (υ − ω) = (εL + εU ). The optimal guard-
banding interval (υ, ω) or the pair (εL , εU ) with the where P (x ∈ [L, U ]) is the probability that the value being
smallest (εL + εU ) can be determined so that β β0 measured lies in [L, U ].
where β0 preset level and the expression for β are given
+∞
ω
by equation (20).
RC = β = h(x, y)dxdy
υ
U
3 Study of customer risk ω L
+ h(x, y)dxdy, (2)
The customer risk is the percentage of non-conforming υ −∞
products that are delivered, and accepted by the customer.
it is calculated as the product of the probability of mak- where h(x, y) is the joint density function of X and Y . It
ing a “non-conforming” product (property of the produc- can easily be shown that X and Y jointly follow a bivariate
tion process) by the (conditional) probability of measuring normal distribution with a mean vector of (μ, μ) and a
“compliant” (i.e. in the tolerance). variance-covariance matrix of Σ given by:
Let εL and εU denote the widths of guardbanding as- 2 2
var (x) cov (x, y) σx σx
sociated with the lower and upper specification limits, re- Σ= = (3)
spectively. For the simplicity of notation, υ = L + εL and cov (x, y) var (y) σx2 σy2
H. Mezouara et al.: Determination of the guardbanding to ensure acceptable risk decision 206-p3
where σy2 represents the variance of the marginal distribu- Can be written as equation (8):
tion of Y , and σy2 = σx2 + σy|x
2
. Note that γ, the correlation
⎡ ω−x ⎤
coefficient of X and Y , is defined as: ⎡ ⎤
∞ ω ∞ σy|x
⎢ ⎥
⎣ g(y|x)dy ⎦ f (x)dx = ⎢ φ(z)dz ⎥
cov X, Y σx ⎣ ⎦ f (x)dx
γ= = σ . (4) U υ U υ−x
∞
ω−x υ−x
= Φ −Φ f (x)dx
The marginal distribution X and Y , assume that the den- σy|x σy|x
sity h(x, y) exists, we denote by g(y|x) and f (x) [6–8]. U
∞ ∞
And ω−x ν −x
= Φ f (x)dx − Φ f (x)dx
h (x, y) = g (y|x) f (x), (5) σy|x σy|x
U U
∞
where ω − (μ + σx λ)
2 = Φ φ(λ)dλ
1 − (ȳ−x)
2
σy|x
2σ
g (y|x) = √ e ȳ|x (6) U −μ
σx
2πσy|x
∞
ν − (μ + σx λ)
and − Φ φ(λ)dλ. (12)
σy|x
2
1 − (x−μ) U−μ/σx
f (x) = √ e 2σx2 . (7)
2πσx
Here, Φ(.) and ϕ(.) represent the cumulative distribution
The first double integral in equation (2) can be written as: and probability density functions of the standard normal
distribution, respectively [9]. Using the following identity,
ω ∞ ∞ ω
h(x, y)dxdy = h(x, y)dydx
υ U U υ ∞
⎡ ⎤ Φ(a + bλ)φ(λ)dλ = BVN √
a
; −K; √
b
∞ ω 2
1+b 1 + b2
= ⎣ g(y|x)dy ⎦ f (x)dx (8) K
(13)
U υ
where BVN (α, β, ρ) represents a function with two vari-
ables standard normal distribution with a correlation co-
where g(y|x) and f (x) are the conditional distribution of efficient of ρ, which is defined by:
Y since X = x and the marginal distribution of X, re-
spectively. Noting that
1
BVN (α, β, ρ) =
2π 1 − ρ2
Ȳ
X=x
2
≈ N x, σy|x
α β 2
x − 2ρxy + y 2
× exp − dxdy
and 2(1 − ρ2 )
−∞ −∞
X ≈ N μ, σx2 . (9) (14)
y−x
Let z = σy|x , and υ y ω is then given
with
υ−x ω−x
z . (10) dΦ (x)
σy|x σy|x ϕ (x) = (15)
dx
h
And
Φ (x) = ϕ (x) dx (16)
x−μ
λ= , −∞
σx
Equation (12) can be simplified to Table 1. Conditions of the test of the proposed model.
⎛ ⎞ 5 ± 2 μm
ω ∞ Horizontal deformation under
ω−μ U −μ
h(x, y)dxdy = BVN ⎝ ;− ; −γ ⎠ Frequency 10 Hz
σ 2 + σ 2 σx Number of pulses 10
υ U x y|x The pulse repetition period 3 ± 0.1 s
⎛ ⎞
Rise time in load 124 ± 4 ms
υ−μ U −μ
− BVN ⎝ ;− ; −γ ⎠, Poisson’s ratio 0.35
σ 2 +σ 2 σx
x y|x
(18) Table 2. Numerical results of the customer risk of the stiffness
modulus.
Similarly, it can be shown that the second integral in equa- εL /εU 0.00 0.25 0.50 0.75 1.00
tion (2) becomes: 0.00 0.70665 0.70669 0.70675 0.70679 0.70684
⎛ ⎞ ⎛ ⎞ 0.25 0.70660 0.70664 0.70669 0.70674 0.70678
ω L
ω − μ υ − μ 0.50 0.70654 0.70654 0.70664 0.70668 0.70673
h(x, y)dxdy = Φ ⎝ ⎠ −Φ ⎝ ⎠ 0.75 0.70649 0.70653 0.70658 0.70663 0.70668
σ 2 +σ 2 σ 2 +σ 2
υ −∞ x y|x x y|x 1.00 0.70644 0.70648 0.70653 0.70658 0.70662
⎛ ⎞
υ − μ L−μ
+ BVN ⎝ ;− ; −γ ⎠ The supplier has taken ten cylindrical specimens (n =
2
σx2 +σy|x σx
10) for the control, the variability of the 10 measures is
⎛ ⎞ given by:
ω − μ L−μ
− BVN ⎝ ;− ; −γ ⎠.
2
σ +σ 2 σx (ni − x)2
x y|x σy|x = = 296 MPa (21)
n−1
(19)
with
Using equations (18) and (19), the customer risk Rc can
be written as: σy2 = σx2 + σȳ|x
2
(22)
ω−μ υ−μ
Rc = β = Φ −Φ σy = σx2 + σȳ|x 2 = σx2 + σ 2y|x = 393.78 MPa (23)
σy σy n
ω−μ U −μ
+ BVN ;− ; −γ and
σy σx cov(X, Y ) σx
γ= = = 0.97. (24)
ω−μ L−μ σȳ
− BVN ;− ; −γ var(X).var(Y )
σy σx
υ−μ U −μ Solving the mathematical model requires a lot of com-
− BVN ;− ; −γ
σy σx puting resources mainly due to the evaluation of bivari-
ate normal probabilities. However, an approximation al-
υ−μ L−μ
+ BVN ;− ; −γ . (20) gorithm, developed with the free software R programming
σy σx language [12], was utilized to evaluate the integrals bivari-
ate normal.
We suppose that β0 = 0.70648% the optimal solution
4 A numerical example to the example problem is found to be υ ∗ = 6000 MPa,
To demonstrate the proposed model, consider an example ω ∗ = 9999 MPa, εL = 1.00 MPa and εU = 0.00 MPa,
of indirect tensile tests of stiffness modulus, according to n = 10, with acustomer risk of 0.70644%, guardband-
standard NF EN 12697-26:2004 [10], tests on cylindrical ing provide a way of assuring that good product would
specimens 100 mm in diameter, with thickness of 53 mm be accepted 99.23%. Numerical results are summarized in
and 2451 kg/m2 density, were carried out under the con- Table 2.
ditions listed in Table 1. ∗
We obtained a Stiffness modulus E = 6696 MPa,
with an estimated
∗ standard deviation (σ) of 382.50 MPa 5 Conclusion
∗ E
with E = 10 i , n = 10 (the average shear modu-
lus corrected). The uncertainty has been calculated by In many industrial settings, manufacturing processes of-
the testing laboratory using an analytical method based ten involve inspection procedures which frequently require
on [11]. The requirements agreed upon between the cus- measurement of a quality characteristic. Various economic
tomer and the supplier specified a lower specification limit, aspects of measurement errors have long been studied.
L, of 6000 MPa, and an upper specification limit, U , Further, there has been a great deal of research efforts
of 10 000 MPa. to reduce the impacts of errors. This paper proposes the
H. Mezouara et al.: Determination of the guardbanding to ensure acceptable risk decision 206-p5
model for economic design of measurement systems by in- 4. J. Chandra, S. Schall, The use of repeated measurements
corporating the concepts of measurement precision and to reduce the effect of measurement errors, IIE Trans. 20,
guardbanding. 83–87 (1988)
The proposed model incorporates various economic as- 5. A. Eagle, A method for handling errors in testing and mea-
pects of measurement errors when determining the level suring, Industrial Quality Control 10(5), 10–15 (1954)
of measurement precision and widths of guardbanding to 6. J. Brucker, A note on the bivariate normal distribution,
minimize the impacts of measurement errors. A numerical Commun. Statist A 8, 175–177 (1979)
example demonstrates the proposed model. 7. Fraser DAS, F. Streit, A further note on the bivariate nor-
mal distribution, Commun. Statist A 9, 1097–1099 (1980)
8. M. Ahsanullah, Some characterizations of the bivariate
normal distribution, Metrika 32, 215–218 (1985)
References 9. S. Chen, K. Chung, Selection of the optimal precision
level and target value for a production process: the lower-
1. D. McCarville, D. Montgomery, Optimal guard bands for specification limit case, IIE Trans. 28, 979–985 (1996)
gauges in series, Quality Eng. 9, 167–177 (1996) 10. AFNOR, Méthodes d’essai pour mélange hydrocarboné à
2. B. Hutchinson, Setting guardband test limits to sat- chaud : Module de Rigidité, NF EN 12697-26 (2004)
isfy MIL-STD-45662A requirements, Proc. of 1991 NCSL 11. AFNOR, Guide pour l’expression de l’incertitude de
Workshop and Symposium (1991), pp. 305–309 mesure, NF ENV 13005 (1999)
3. D. Deaver, Using Guardbands to Justify TURs Less than 12. P.Dalgaard, Introductory Statistics with R, 2nd edn.
4:1, Proc. of ASQC 49th Annual Quality Congress (1995), (Springer, New York, 2008)
pp. 136–141